A target scene spectral library construction method of frequency domain structured modeling and spectral unmixing

Abstract

The present application relates to the technical field of hyperspectral remote sensing image processing and spectral information modeling, and particularly relates to a target scene spectral library construction method of frequency domain structured modeling and spectral unmixing. First, spatial structure modeling is performed on the observation image of the target scene to generate a hyperspectral intermediate representation that maintains spatial structure consistency; then, the hyperspectral intermediate representation is corrected for spectral consistency under the guidance of information constraints to suppress unstable disturbances across spectral channels; second, mixed spectral signals in the target scene are separated through spectral unmixing to extract representative spectral components; finally, the representative spectral components are screened and organized to construct a target scene spectral library and output as spectral priori. The present application realizes automatic generation of target scene spectral priori with physical meaning and stability from low-dimensional observation images, and is suitable for hyperspectral reconstruction, target recognition and interpretation and other remote sensing downstream applications that rely on spectral priori information.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing image processing and spectral information modeling technology, specifically to a method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing. Background Technology

[0002] Hyperspectral remote sensing imagery can simultaneously acquire spatial information and continuous, detailed spectral information about ground features, making it valuable for applications in areas such as feature identification, target detection, environmental monitoring, and quantitative analysis. However, the complex structure, high cost, and limited acquisition efficiency of hyperspectral imaging systems make it difficult to achieve large-scale, real-time acquisition of hyperspectral data in practical applications, which severely restricts its widespread application in various downstream tasks.

[0003] To alleviate these problems, researchers have proposed various methods such as spectral super-resolution, hyperspectral reconstruction, and cross-modal mapping, aiming to recover high-dimensional spectral information from low-dimensional observation images (such as RGB or multispectral images). These existing methods mainly achieve spectral estimation by establishing a mapping model between observation images and hyperspectral images. However, since the information provided by low-dimensional observations has limited constraints on hyperspectral information, this type of problem is inherently ill-conditioned, meaning that the reconstruction results are highly dependent on the model structure and data distribution characteristics, and are prone to problems such as spectral instability or spatial structure distortion.

[0004] Furthermore, there is relatively little research on the construction of spectral priors for target scenes in existing methods. Related technologies often focus on a single reconstruction result or end-to-end mapping process, failing to systematically and collaboratively model the spatial structural features and intrinsic spectral composition of the target scene. This results in insufficient representativeness, stability, and reusability of the obtained spectral components, making it difficult to provide reliable prior support for subsequent downstream tasks such as spectral reconstruction, classification, or interpretation.

[0005] Therefore, there is an urgent need for a method that can start from low-dimensional observation images of the target scene, jointly model the spatial structure information and spectral composition characteristics of the scene, analyze and screen the inherent spectral components of the target scene, and thus automatically generate a physically meaningful, stable and representative spectral prior of the target scene. Summary of the Invention

[0006] This invention addresses the problem that existing methods for acquiring hyperspectral data of target scenes are limited, making it difficult to provide stable and reliable spectral priors for downstream tasks such as spectral reconstruction, classification, or interpretation. It proposes a method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing. The core of this method is the automatic construction and stable representation of the target scene spectral prior: First, spatial structure modeling is performed on the observed image of the target scene to generate a hyperspectral intermediate representation that maintains spatial structure consistency. Then, under the constraint of guiding information, the spectral consistency of the hyperspectral intermediate representation is corrected to suppress unstable perturbations across spectral channels. Next, the mixed spectral signals in the target scene are separated through spectral demixing to extract representative spectral components. Finally, the representative spectral components are screened and organized to construct the target scene spectral library, which serves as the spectral prior output. This invention enables the automatic generation of physically meaningful and stable target scene spectral priors from low-dimensional observed images, making it particularly suitable for downstream remote sensing applications such as hyperspectral reconstruction, target identification, and interpretation that rely on spectral prior information.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing includes the following steps:

[0009] Step 1: Construct a target scene spatial structure reconstruction network and generate a hyperspectral intermediate representation.

[0010] Using the observed images of the target scene as input, a spatial structure reconstruction network is constructed to model the spatial information of the target scene and output the hyperspectral intermediate representation of the target scene, which is used to describe the high-dimensional spectral features of the target scene that remain consistent in the spatial dimension.

[0011] Furthermore, step 1 includes the following sub-steps:

[0012] Step 1.1: Let the observed image of the target scene be represented as... K, H, and W represent the number of bands and size of the observed image; the wavelet spatial structure decomposition operator can be used. The observed images are processed to obtain several spatial structure feature components. Each spatial structural feature component represents the structural information of different scales or different spatial frequency characteristics in the target scene, and is used to separate low-frequency structural information and high-frequency detail information in the target scene.

[0013] Step 1.2: For the spatial structural feature components, construct feature modeling functions for each component, perform global correlation modeling on each component, and introduce collaborative constraints between components to ensure that each spatial feature satisfies spatial consistency requirements while maintaining its own structural characteristics; map the collaboratively modeled spatial features to the hyperspectral dimension to obtain the hyperspectral intermediate representation of the target scene. ,in .

[0014] Step 2: Construct a spectral consistency correction network and obtain a spectrally stable hyperspectral representation.

[0015] The hyperspectral intermediates output from step 1 are represented as follows. Using the input as input, a spectral consistency correction network is constructed to constrain and adjust the hyperspectral intermediate representation in the spectral dimension, and output a hyperspectral representation with continuous and stable spectral response.

[0016] Furthermore, step 2 includes the following sub-steps:

[0017] Step 2.1: Based on the observed image of the target scene or its derived features, extract the semantic prior information of the target scene, and map the semantic prior information into guiding weights that modulate spatial location. It is used to adjust the spectral correction amplitude and indicate the relative importance of different spatial regions in the spectral consistency correction process.

[0018] Step 2.2: Under the constraint of the guiding weight, perform hyperspectral intermediate representation Spectral consistency correction is performed to adjust the fluctuation components across spectral channels, resulting in a hyperspectral representation with enhanced spectral consistency. .

[0019] Step 3: Perform spectral unmixing and extract representative spectral components of the target scene.

[0020] The hyperspectral representation obtained in step 2 As input, a spectral demixing module is constructed to separate the mixed spectral signals in the target scene, obtaining a set of candidate spectral components. and their corresponding spatial distribution weights; introduce reconstruction consistency constraints on the candidate spectral components to ensure they satisfy the approximate reconstruction relationship. .

[0021] Step 4: Construct a spectral library for the target scene and output the spectral prior results.

[0022] The endmember spectrum set obtained in step 3 The endmember spectra are screened according to a preset spectral similarity threshold to remove redundant or unstable spectral components. The screened endmember spectra are then structured and associated with each other to generate a target scene spectral library that represents the representative spectral features of the target scene. This target scene spectral library is then used as a spectral prior output to support subsequent downstream application tasks such as spectral reconstruction, classification, or interpretation.

[0023] Compared with existing methods, the advantages of the present invention are:

[0024] Taking into full account the high coupling between spatial structure and spectral information in hyperspectral remote sensing images, this study employs a collaborative design of frequency-domain structured modeling and spectral unmixing to enable the model to possess more stable feature representation capabilities across target scenes with different spatial scales and structural morphologies. By introducing a learnable wavelet spatial structure decomposition operator, the multi-scale spatial structure of the target scene is decomposed and reconstructed in the frequency domain, effectively enhancing the joint modeling capability of low-frequency overall structure and high-frequency detail information, and improving the accuracy of spatial structure preservation in complex scenes. Based on spatial structure consistency constraints, a semantic prior-guided spectral consistency correction mechanism is combined to suppress unstable perturbations across spectral channels. Furthermore, spectral unmixing fully explores the spectral composition characteristics within the target scene, achieving effective separation and stable representation of spectral information. Finally, a method for automatically constructing a target scene spectral library for downstream tasks is proposed, capable of generating physically meaningful, representative, and reusable spectral priors from low-dimensional observation images, providing reliable support for remote sensing applications such as hyperspectral reconstruction, classification, and interpretation. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing according to the present invention;

[0026] Figure 2 This is a schematic diagram of the target scene spatial structure reconstruction network constructed in this invention.

[0027] Figure 3 This is a schematic diagram of the spectral consistency correction network constructed in this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0029] like Figure 1 As shown, a method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing includes the following steps:

[0030] Step 1: Construct the target scene spatial structure reconstruction network.

[0031] Using the observed image of the target scene as input, the spatial structure of the observed image is modeled and reconstructed, spatial information with different structural characteristics in the target scene is extracted, and a hyperspectral intermediate representation that maintains the consistency of spatial structure is generated.

[0032] Specifically, step 1 includes the following steps:

[0033] Step 1.1: Using the observed image of the target scene as input, a learnable wavelet decomposition operator with orthogonality and energy conservation constraints is introduced to perform frequency domain structural decomposition on the observed image, resulting in multiple structural components with different spatial frequency characteristics. The low-frequency structural components are used to characterize the overall spatial layout and regional structural information in the target scene, while the high-frequency structural components are used to characterize the edge, texture, and local detail changes in the target scene. Each structural component shares a unified wavelet constraint condition during the decomposition process to ensure the consistency and stability of the spatial structural decomposition.

[0034] Step 1.2: For the different frequency structural components obtained in Step 1.1, construct corresponding spatial structural modeling units, map each structural component into a sequence representation, and use the Transformer modeling module based on multi-head self-attention mechanism to model them, so as to capture the global dependencies between spatial locations. The low-frequency structural components adopt a complete global self-attention modeling method, and the high-frequency structural components adopt a restricted update modeling method, so that the high-frequency structure can be locally adjusted while maintaining the original spatial form.

[0035] Step 1.3: After completing the spatial modeling of each frequency structural component, a spatial cooperative constraint mechanism is introduced to coordinate the spatial responses between different structural components, ensuring that they maintain a consistent structural correspondence in spatial location. The structural components after cooperative constraint are mapped back to the spatial domain using the wavelet inverse transform operator and uniformly mapped to the target hyperspectral dimension. Furthermore, based on the spatial cooperative constraint, a continuity constraint is applied to the response variation amplitude between spatially adjacent pixel positions and spectrally adjacent channels to suppress local instability fluctuations that may be introduced during high-frequency modeling. A hyperspectral intermediate representation of the target scene that maintains spatial structural consistency is generated.

[0036] The constructed target scene spatial structure reconstruction network, such as Figure 2 As shown.

[0037] Step 2: Construct a spectral consistency correction network.

[0038] Using the hyperspectral intermediate representation obtained in step 1 as input, a spectral consistency correction network is constructed. While keeping the spatial structure unchanged, the spectral dimensions are adjusted to correct the inconsistencies between cross-spectral channels, resulting in a hyperspectral representation of the target scene with enhanced spectral consistency.

[0039] Step 2 includes the following steps:

[0040] Step 2.1: Based on the observed images of the target scene, construct a semantic prior extraction function to extract prior representations that reflect the semantic attributes of the target scene; and map the semantic prior representations into spatially related guiding weights to characterize the modulation intensity of different spatial regions during the spectral correction process.

[0041] Step 2.2: Using the hyperspectral intermediate representation obtained in Step 1.3 As input, it is reorganized into a spectral sequence along the spectral dimension, and a spectral sequence modeling module based on the Transformer structure is constructed to model the correlation between cross-spectral channels and obtain the spectral residual signal. Under the guiding weight constraint constructed in step 2.1, the spectral residual signal is subjected to amplitude constraint and weighted modulation to correct the non-uniform fluctuations across spectral channels and output a hyperspectral representation of the target scene with enhanced spectral consistency.

[0042] The constructed spectral consistency correction network is as follows: Figure 3 As shown.

[0043] Step 3: Perform spectral unmixing and extract spectral components.

[0044] Using the hyperspectral representation obtained in step 2 as input, a spectral demixing module is constructed to separate the mixed spectral signals in the target scene, extract candidate spectral components and their corresponding spatial distribution relationships, and obtain a set of representative spectral components of the target scene through constraints.

[0045] Step 4: Construct a spectral library for the target scene.

[0046] The representative spectral components obtained in step 3 are screened and organized to construct a target scene spectral library, which is then used as a spectral prior output to support subsequent hyperspectral reconstruction and related downstream tasks.

[0047] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.

Claims

1. A method for constructing a target scene spectral library using frequency domain structured modeling and spectral demixing, characterized in that, First, spatial structure modeling is performed on the observed images of the target scene to generate a hyperspectral intermediate representation that maintains spatial structure consistency. Then, under the constraint of guiding information, the hyperspectral intermediate representation is corrected for spectral consistency to suppress unstable disturbances in cross-spectral channels. Next, the mixed spectral signals in the target scene are separated by spectral demixing to extract representative spectral components. Finally, the representative spectral components are screened and organized to construct a spectral library of the target scene and serve as a spectral prior output.

2. The method for constructing a target scene spectral library through frequency domain structured modeling and spectral demixing according to claim 1, characterized in that, Specifically, the following steps are included: Step 1: Construct the target scene spatial structure reconstruction network; Observation images of the target scene The image is input into a spatial structure reconstruction network, where a learnable wavelet spatial decomposition operator is used to perform multi-scale spatial-frequency domain decomposition on the observed image, obtaining spatial feature components of multiple different frequency sub-bands. Feature extraction and reconstruction processing are then performed on each of these spatial feature components to enhance the stability and detail consistency of the target scene's spatial structure, outputting a hyperspectral intermediate representation of the target scene. ,in ; Step 2: Construct a spectral consistency correction network and perform spectral correction; The hyperspectral intermediate representation obtained in step 1 The input is fed into the spectral consistency correction network, which generates guiding weights to adjust the spectral correction amplitude based on the semantic categories or structural features of different spatial regions in the target scene. Under the constraint of the guiding weight, the hyperspectral intermediate representation The system performs serialization along the spectral dimension and uses a residual network to suppress inconsistencies between spectral channels, thereby reducing cross-channel noise and redundant information, resulting in a hyperspectral representation corrected for spectral consistency. ; Step 3: Represent the hyperspectral data obtained in Step 2 The input is fed to the spectral demixing module, which decomposes the mixed spectral signals in the hyperspectral representation to obtain the endmember spectral set of the target scene. and the corresponding spatial abundance distribution and the corresponding abundance distribution Each end element During the unmixing process, a reconstruction consistency constraint is introduced to ensure that the hyperspectral representation Xr satisfies an approximate reconstruction relation: ; Step 4: Based on the endmember spectrum set obtained in Step 3 The spectra of each endmember are filtered based on spectral similarity thresholds to remove redundant or unstable spectral components, generating a target scene spectral library to characterize the representative spectral features of the target scene. The target scene spectral library Used to provide spectral prior information for subsequent spectral reconstruction, classification, or interpretation tasks.

3. The method for constructing a target scene spectral library through frequency domain structured modeling and spectral demixing according to claim 2, characterized in that: Step 1 includes the following sub-steps: Step 1.1: Perform multi-scale spatial decomposition on the observed image of the target scene using a learnable wavelet spatial structure decomposition operator, mapping the observed image into multiple sets of spatial feature components with different spatial resolutions and structural sensitivities. To separate low-frequency structural information from high-frequency detail information in the target scene; Step 1.2: For each of the spatial feature components, construct corresponding feature modeling units, model the correlation between different spatial locations through a multi-head self-attention mechanism, and introduce collaborative constraints to limit the deviation of different feature components in spatial structure, thereby obtaining sub-component feature representations with consistent spatial structure. Step 1.3: Apply amplitude constraints to the sub-component feature representation to limit the variation amplitude between spatially adjacent pixel positions and spectrally adjacent channels. While maintaining the continuity of the spatial structure, map the sub-component features to the hyperspectral dimension of the target scene and output the spatially reconstructed hyperspectral intermediate representation. .

4. The method for constructing a target scene spectral library through frequency domain structured modeling and spectral demixing according to claim 2, characterized in that: Step 2 includes the following sub-steps: Step 2.1: Based on the observed image corresponding to the target scene, obtain the semantic prior representation reflecting the spatial region category or structural attribute of the target scene through the semantic prior extraction module, and map the semantic prior representation into guiding weights that have a modulation effect on the spectral channels of different spatial locations, which are used to indicate the relative importance of each region in the spectral correction process. Step 2.2: Under the constraint of the guiding weight, convert the hyperspectral intermediate representation output in Step 1... The input is fed into the residual network to suppress non-uniform disturbances between spectral channels, reduce redundant information and noise components across spectral channels, and enable the output hyperspectral representation to have higher spectral smoothness and consistency while maintaining the spatial structure.

5. The method for constructing a target scene spectral library through frequency domain structured modeling and spectral demixing according to claim 2, characterized in that: Step 3 includes the following sub-steps: Step 3.1: Represent the hyperspectral image after spectral consistency correction. As input, the mixed spectral signals in the target scene are decomposed using a spectral demixing operator to obtain a set of candidate endmember spectra. During the decomposition process, a spectral reconstruction consistency constraint is introduced to ensure that the spectra of each endmember and their corresponding spatial abundance distributions satisfy the requirements for hyperspectral representation. The approximate reconstruction relationship is obtained, and the candidate endmember spectra are screened and integrated based on spectral similarity, stability or contribution index.

6. The method for constructing a target scene spectral library through frequency domain structured modeling and spectral demixing according to claim 2, characterized in that: Step 4 includes the following sub-steps: Step 4.1: Organize the endmember spectrum set obtained in Step 3 in a structured manner, perform correlation modeling on the endmember spectra according to spectral similarity or spatial co-occurrence relationships, form a representative and stable target scene spectral library, and use the target scene spectral library as the spectral prior output.

Images